Self-reported inner speech relates to phonological retrieval ability in people with aphasia

Abstract

Many individuals with aphasia report the ability to say words in their heads despite spoken naming difficulty. Here, we examined individual differences in the experience of inner speech (IS) in participants with aphasia to test the hypotheses that self-reported IS reflects intact phonological retrieval and that articulatory output processing is not essential to IS. Participants (n=53) reported their ability to name items correctly internally during a silent picture-naming task. We compared this measure of self-reported IS to spoken picture naming and a battery of tasks measuring the underlying processes required for naming (i.e., phonological retrieval and output processing). Results from three separate analyses of these measures indicate that self- reported IS relates to phonological retrieval and that speech output processes are not a necessary component of IS. We suggest that self-reported IS may be a clinically valuable measure that could assist in clinical decision-making regarding anomia diagnosis and treatment.

1. Introduction

The hallmark deficit of aphasia is anomia, a deficit in naming and word-finding (Goodglass & Wingfield, 1997; Laine & Martin, 2006). Interestingly, many individuals with aphasia make comments suggesting that their spoken output does not fully reflect their lexical knowledge, e.g., “My head is saying it, but my mouth is not saying it at all!” This reported discrepancy between inner and overt speech has been confirmed in some prior research in individuals with aphasia (Feinberg, Rothi, & Heilman, 1986; Geva, Bennett, Warburton, & Patterson, 2011; Hayward, Snider, Luta, Friedman, & Turkeltaub, 2016; Stark, Geva, & Warburton, 2017). Understanding the relationship of self-reported inner speech to objective measures of word retrieval and speech production may reveal the significance of this experience in aphasia and elucidate the nature of self-perceived inner speech more broadly. Here, we examine a group of people with aphasia to identify objective language abilities that are associated with more frequent subjective experience of successful inner speech (successful IS).

IS is defined most simply as the imagery of having an inner voice in one’s head (Sokolov, 1972). The phenomenon has been studied scientifically for over a century, with researchers identifying the crucial role of IS in various cognitive processes, including language learning, reading, working memory, decision-making, and even self-awareness (Alderson-Day & Fernyhough, 2015; Baddeley & Hitch, 1974; Perrone-Bertolotti, Rapin, Lachaux, Baciu, & L??venbruck, 2014; Vygotsky, 1962). Early theories proposed that IS requires the same articulatory movements as overt speech, only to a lesser degree (Watson, 1913). Such strong theories have since been dispelled (e.g., by Smith, Brown, Toman, & Goodman, 1947), but the idea that IS requires a fully specified articulatory plan persisted in some processing models (Levelt, 1983; Postma & Noordanus, 1996). Similarly, recent neuroimaging studies have suggested that IS is generated and monitored on the basis of efference copies from motor systems (Tian & Poeppel, 2013, 2015). In contrast, IS has also been defined as more abstract in nature, without need for specific articulatory features (Indefrey & Levelt, 2004; Levelt, 2001; Oppenheim & Dell, 2008). Prior studies also support an intermediate stance, in which IS is a flexible construct that can evoke speech production processes under certain circumstances (Oppenheim & Dell, 2010; Sokolov, 1972). Given these disparate findings and models, it remains unclear if articulatory processes are a necessary component of IS or if earlier stages of production, such as lexical phonology, are available to conscious monitoring and are thus sufficient for an experience of IS.

This question has been addressed to some degree using motor interference tasks during IS in healthy language users, although findings have been inconsistent. For example, articulatory suppression has been shown to impair accuracy on phoneme monitoring when the task is based on written words (Smith, Reisberg, & Wilson, 1992) but not when based on internally, silently generated words (Wheeldon & Levelt, 1995). Another set of studies has demonstrated that articulatory interference (generated via a distracting, overt speech task) impairs performance on a variety of silent problem-solving tasks, but that the effects are modulated by task difficulty and the nature of the relationship between the target and distractor tasks (Sokolov, 1972). Overall, these interference approaches in healthy language users provide mixed results as to the role of articulation in IS. Even in the case of studies that do find a significant impact of motor interference, it remains unclear whether subarticulation is relevant to IS itself or to the judgment required by the specific task. Examining people with post-stroke aphasia provides a powerful way to address this issue because word retrieval and speech production can be dissociable in this population (Goodglass & Wingfield, 1997; Laine & Martin, 2006). By examining the relationship of IS to deficits in word retrieval vs. speech production, we can learn whether earlier stages of speech processing are available to conscious self-monitoring and may be perceived as IS in certain situations.

The majority of previous research on IS in individuals with aphasia has used performance on objective language tasks (e.g., silent rhyme or homophone judgments on written word pairs) as a proxy for the preservation of IS (Feinberg et al., 1986; Geva et al., 2011; Langland-Hassan, Faries, Richardson, & Dietz, 2015; Stark et al., 2017). With this approach, findings have consistently shown that IS can be preserved relative to overt speech in individuals with aphasia, particularly in individuals with conduction aphasia or verbal apraxia (a motor speech disorder), two diagnoses associated with deficits in speech output processing. Importantly, these studies have not consistently matched the specific stimuli and/or task structure across the inner/overt speech tasks, so it is not yet known whether these discrepancies would be observed in carefully matched tasks. Furthermore, studies relying on objective measures as a proxy for IS ability overlook what may be an important potential source of information about the phenomenon of IS: the insight of the individual language user who is experiencing it.

Self-reports have been used in research on IS in healthy language users (Hurlburt, Alderson-Day, Kühn, & Fernyhough, 2016; Morin, Uttl, & Hamper, 2011) as well as in prior work on a related phenomenon, the tip-of-the-tongue experience, in individuals with aphasia (Goodglass, Kaplan, Weintraub, & Ackerman, 1976). In our own prior studies, we have assessed IS based on the subjective insight of the individual, then examined relationships between these insights and objective language scores in order to better understand the nature of IS. We have hypothesized that the subjective experience of successful IS arises in conjunction with successful lexical phonological retrieval and that subsequent processing stages required for speech output are not required for an internal sense of successful IS.

Our previous studies have provided initial support for these hypotheses. We found significant relationships between item-level reports of successful IS during silent picture-naming and accuracy/error type on a spoken naming task in a small group of individuals with aphasia (Hayward, 2016; Hayward et al., 2016). Using similar methods, we then replicated these findings in a larger participant group, again showing that self-reported IS is closely related to lexical retrieval (Fama et al., 2019). These same studies also revealed preliminary evidence that IS does not relate to post-lexical output processing. Self-reported IS related to word features primarily associated with retrieval, but did not relate to the articulatory complexity of words (Fama et al., 2019; Hayward, 2016). In a separate line of research using an interview-based approach, we found that the experience of successful IS in the specific context of anomia (where successful IS is followed by a failure of spoken output) relates to phonological output processing and to lesions in left ventral sensorimotor cortex (Fama, Hayward, Snider, Friedman, & Turkeltaub, 2017). Taken together, these findings suggest that IS is closely related to word retrieval and that articulatory processing may not be an essential component of IS itself.

While these previous studies have provided promising evidence for our hypotheses, the scope of the conclusions drawn have been relatively limited, due to small number of participants (Hayward, 2016; Hayward et al., 2016), large proportion of participants excluded from main analysis (Fama et al., 2019), or examination of successful IS only in the specific context where it precedes anomia (Fama et al., 2017). In the current study, we approached the study of self- reported IS at the level of overall task performance, which allowed us to examine relationships between the experience of successful IS during silent picture naming and both lexical retrieval and speech production ability. Specifically, we compared the number of items for which participants reported successful IS during a silent-picture naming task to performance on a battery of language tests designed to assess lexical retrieval and speech output processing. In addition, we compared reports of successful IS to the types of errors made on a spoken naming task, since deficits at various stages of the mental process of naming tend to cause particular error patterns (e.g., phonological errors tend to occur due to failure at the level of phonological retrieval or in post-lexical output processing). Based on our hypotheses, we expected that the rate at which people report successful IS should relate to their phonological retrieval ability, but not their speech production ability. Further, we predicted that participants reporting more frequent successful IS would produce naming errors that suggest relatively intact phonological retrieval.

2. Methods

2.1. Participants

Two participant groups were utilized for this study: adults in the chronic stage of recovery from left-sided stroke and healthy, age-matched subjects for task norming. All participants underwent an informd consent process approved by the Georgetown University Institutional Review Board.

Patient participants for this study were 53 adults, all native English speakers, who suffered a left-sided stroke at least six months prior to enrollment. Several participants had evidence of prior small, incidental strokes that were asymptomatic: one in the right putamen, two in the right cerebellum, one in the left cerebellum, and two in right hemisphere cortical areas. 65 participants were initially enrolled, but two were unable to complete both sessions of the language testing, and nine failed to demonstrate adequate auditory comprehension for inclusion, defined by a score of at least 48/60 on the “Yes/No Questions” subtest of the Western Aphasia Battery – Revised) (Kertesz, 2006). We also excluded one participant who demonstrated near floor performance on all tasks in the battery (despite meeting the comprehension cut-off). The final participant group thus included 53 participants (22 women), with a mean age of 60.2 years (SD=9.8, range 40–80), mean education of 16 years (SD=2.8, range 12–24), and mean time since stroke of 5 years (SD=4.8, range 0.5–22.9), with handedness as follows: 46 right-handed, 6 left- handed, and one ambidextrous. All participants in this final group presented with adequate single-word intelligibility, i.e., no participants had significant evidence of dysarthria, as assessed by two certified speech-language pathologists (authors MEF and SFS).

Participants in the norming sample were a set of 20 healthy older adults who were native English speakers and reported no history of developmental learning disability, neurological disorder, or major psychiatric illness. In this group, participants were an average age of 65.7 years old (SD=8.4) with an average education of 17.2 years (SD=2.6).

2.2. Session structure, stimuli, and norming

Participants completed the test battery across two sessions occurring at least 10 days apart (mean 18.7 days). The primary stimuli were 60 items, including 20 each of 1-, 2- and 3- syllable words. For IS Report and Spoken Naming (see below), we included an additional 60 items from the Philadelphia Naming Test (PNT), for a total of 120 items on each of these tests (Roach, Schwartz, Martin, Grewal, & Brecher, 1996). For First Letter Identification and Syllable Counting (see below), the 60-item list was split into two matched sets for the picture-based and auditory versions of the tasks. All tests were normed in healthy, older adults and any specific items or trials missed by more than 25% of participants in the norming sample were removed from analysis. For a complete description of the stimuli used for this study, including psycholinguistic features as well as sources of picture stimuli, please see Fama et al. (2019).

All tasks requiring spoken responses were video recorded for offline scoring and the first complete attempt for each trial was scored. Two raters independently scored the spoken naming task (authors MEF and MPH) and the motor speech evaluation (authors MEF and SFS). For these tasks, discrepancies between the raters were resolved via mutual re-review, and a third rater resolved any further disagreement.

2.3. Language testing battery

2.3.1. IS Report

Items were presented one at a time on a laptop screen and participants were instructed to name the picture in their heads without moving their lips or tongue. They then pressed a button on the keyboard (labeled with the written words “yes” and “no”) to report whether they could say the word in their head, with all the right sounds in the right order. The test items advanced automatically upon key press. The 120 stimulus pictures for the IS Report task were split into two sets of 60 items (Sets 1 and 2), each including 30 in-house items and 30 PNT items. These sets were matched on relevant psycholinguistic variables, including frequency, age of acquisition, length, and articulatory complexity. Each set was administered on a separate day of testing, counterbalanced for order across participants. This task was administered using PsychoPy presentation software (Peirce, 2009), as were all other tasks in the battery with the exception of repetition. For this task and all other tasks where participants were asked not to move their lips or tongue, occasional non-compliance resulted in review of the instructions by the experimenter; no trials were excluded.

2.3.2. Spoken Naming

For the spoken naming task, participants named a set of stimulus pictures that were presented one at a time for up to 20 seconds on a laptop screen. Participants were given instructions to “please use only one word.” Participants advanced the test items by pressing the space bar. The same two 60-item sets of stimuli used for IS Report were also used for Spoken Naming, for a total of 120 items. If set 1 was used for IS Report on day one of testing, set 2 was used for Spoken Naming, and vice versa for day two, such that each item was encountered only once per day. The order was counterbalanced across participants.

2.3.3. Error coding for the spoken naming task

Responses on the spoken naming task were transcribed and scored for accuracy and error type. Two error categories are of interest to our analysis:

• Semantic error: a real word (usually a noun) that is related to the target as a synonym, category coordinate, superordinate, subordinate, diminutive, or associated concept

• Phonological error: a real word or non-word with at least 50% overlap with the phonemes in the target word

Other error types (including unrelated errors, circumlocutions, abstruse neologisms, and “no response” errors) are not examined in the context of this study as we did not have any specific hypotheses regarding these error types.

2.3.4. Sentence-level auditory comprehension

The “Yes/No Questions” subtest of the Western Aphasia Battery – Revised (Kertesz, 2006) was administered to confirm adequate comprehension for the tasks of interest (cut-off score of at least 48/60).

2.3.5. Measures of lexical retrieval

To assess lexical retrieval, we utilized three sets of matched tasks that each included a picture-based version and an auditory version. Each picture-based task required the participant to retrieve the name of the picture in order to perform the task, whereas in the auditory tasks, the word was presented aloud to the participant.

• Rhyme judgment – picture-based

  Participants completed a silent rhyme judgment task in which two pictures appeared on the screen and they were asked to answer (yes/no) whether the names of the pictures rhymed (40 trials total). All items were 1-syllable words. The stimuli for this task were not drawn from the main 60-item list due to constraints for single syllable, rhyming word pairs. Two trials were removed after norming, so only the 38 remaining trials were included in the analyses.

• Rhyme judgment – matched auditory task

  Participants heard two pre-recorded words presented via high-fidelity headphones and were asked to answer (yes/no) whether the two words rhymed. The same items were used as in the picture-based rhyme judgment task, and the same 38 item-pairs were included the analyses.

• First letter identification and syllable counting – picture-based

  In this task, participants were presented with 30 picture stimuli, one at a time, and were asked to name the picture in their heads, without moving their lips or tongue, and then indicate the number of syllables and the first letter of the word by pointing to a response page with the numbers 1–5 and the alphabet (in order, in lowercase Arial font). If participants self-corrected spontaneously, the final answer was accepted. Since these responses were provided by participants via pointing to a response page (vs. answering verbally), we allowed self-correction in order to avoid any negative impact of difficulty with visual grapheme identification or motor control.

• First letter identification and syllable counting – auditory

  In this task, participants were presented with 30 different, matched stimuli and an auditory recording of the target word accompanied the pictures (simultaneous presentation). On the same response page as described above, participants pointed to the number of syllables and the first letter of the word. They were allowed to repeat the word aloud to themselves if they did so spontaneously. As with the picture-based version, spontaneous self-correction was allowed.

The rhyme judgment tasks utilized the exact same items, so these tasks were presented on separate testing sessions. The picture-based task was administered during session one to avoid participants’ potential memory of the word pairs from the auditory task having an impact on performance of the silent task. For the first letter identification and syllable counting tasks, the auditory and picture-based versions used different, precisely matched stimuli, so these tasks were administered on the same day (session two). We administered them back-to-back due to the relative complexity of providing task instructions and demonstration of how to use the response sheet; all participants completed the picture-based task followed by the auditory task.

To isolate the portion of task performance specifically related to phonological retrieval, we subtracted scores on the auditory version of each task, in which the word was provided to the participant and they needed only to make the judgments necessary to perform the task, from the scores on the picture-based task, in which both phonological retrieval and judgments were necessary. A higher picture-based – auditory difference score (in most cases, less negative) represents better phonological retrieval ability.

2.3.6. Measures of speech output processing

To assess post-lexical output processing, we included several tasks that rely on speech output processes without requiring lexical retrieval. We used real words in the context of a general motor speechevaluation, but utilized pseudowords during separate repetition and oral reading tasks in order to limit the support of lexical forms during production.

• Pseudoword repetition

  The repetition task was performed using pre-recorded stimuli in a natural speaker’s voice, played through QuickTime software on a laptop computer. For this task, pronounceable non-words were generated from the 60 real word stimuli by changing at least one phoneme for 1- and 2-syllable words and at least two phonemes for 3-syllable words. Each pseudoword was matched to the corresponding real word for number of syllables and articulatory complexity, measured using the Word Complexity Measure (Stoel-Gammon, 2010). Each item was presented once with a 5-second inter-trial interval, although additional time was provided if necessary. Two items were removed based on norming, so the analysis was conducted on the remaining 58 items.

• Pseudoword oral reading

  Participants completed a 20-item pseudoword reading task. One pseudoword appeared on the laptop screen at a time and participants had 10 seconds to read the pseudoword aloud. All items were single-syllable pseudowords composed of 3–4 letters/phonemes. Accuracy was scored based on participants’ first complete attempt at reading the word, with plausible alternatives allowed based on prior norming.

• Motor speech evaluation

  We used a standard motor speech examination protocol to test speech production ability (Haley, Jacks, de Riesthal, Abou-Khalil, & Roth, 2012). In our analysis, we included the following scores, which measure motor agility, sequencing, phonological processing, and motor speech production:

  • AMR rate: the mean number of syllables produced per second, averaged across separate attempts for /pʌ/, /tʌ/, and /kʌ/
  • Adequate SMR performance: a binary measure reflecting whether or not theparticipant produced at least three accurate syllable triads (/pʌtʌkʌ/ x 3)
  • Segmental errors in multisyllabic word repetition: substitution, omission, or addition of one or more phonemes
  • Ambiguous/distorted consonant productions in multisyllabic word repetition. Two of the above measures, difficulty with sequential motion rates (SMRs) and the production of distorted/ambiguous consonant sounds, are particularly characteristic of apraxia of speech (Jacks & Haley, 2015).

2.4. Statistical analyses

All statistical analyses were performed on the data from patient participants only; noformal analyses were performed on the norming sample. We first calculated overall mean and SD for task performance in the participant group. We performed a dependent-samples t-test to compare performance on Spoken Naming vs. IS Report. We compared overall IS Report scores to (1) age and (2) chronicity using non-parametric Spearman correlations in the entire participant group. Then, we performed a factor analysis to confirm our predictions about the task demands of lexical retrieval vs. output processing. The tests entered into the analysis were: pseudoword repetition, pseudoword reading, the four measures from the motor speech evaluation, and the three difference scores from the matched picture-based/auditory tasks. We implemented a principal components analysis of the correlation matrices, using a standard eigenvalue > 1 cutoff to extract components and Varimax rotation with Kaiser normalization to generate orthogonal factors. The rotated component matrix provided the loadings of each test onto the resulting factors. Factor scores for each subject were determined using the regression method.

The self-reported IS scores (IS Report) were left-skewed, so we applied a logit transform (ln(p/(1-p)) to these scores. Since we used difference scores of the three matched auditory/picture-based tasks (first letter identification, syllable counting, and rhyme judgment) to represent phonological retrieval, we confirmed that logit-transformed IS Report correlated with the picture-based versions of the task using bivariate Pearson’s correlations as well as partial correlations controlling for the auditory version of the task. Using additional bivariate Pearson’s correlations, we then examined relationships between the logit-transformed IS Report scores and spoken naming accuracy (Spoken Naming) and two sets of measures: the factor scores resulting from the principal components analysis and the relative proportions of phonological and semantic errors on the spoken naming test. We set our significance threshold at a Bonferroni- corrected p-value of p < .05 (corrected for nine bivariate correlations performed). In a final analysis, we used hierarchical linear regression to examine the relative contributions of the two factor scores and the IS Report score in predicting performance on Spoken Naming, to determine whether IS Report added significant predictive value beyond the objective measures of retrieval and production. All statistical analyses were performed in SPSS 24.

3. Results

3.1. Overall performance on the language battery

Participants as a group reported successful IS that exceeded Spoken Naming accuracy (t(52) = 8.62, p < 0.001; Cohen’s d = 1.18) (Table 1). This difference is expected if IS reflects an earlier stage of production than subarticulation, i.e., if successful IS requires fewer processing stages than spoken naming does, or if participants are over-reporting the success of IS (see Discussion, section 4.3 for a discussion of potential bias in reporting). Neither age nor stroke chronicity related to IS Report (IS Report vs. age: r_s = −.135, p = .28; IS Report vs. chronicity: r_s= .11, p = .41). As expected, performance on the picture-based retrieval tasks (rhyme judgment, first letter identification, and syllable counting) was lower than performance on each of the auditory matched tasks, since the former require lexical retrieval in addition to phonological judgments.

Table 1. Average performance on the language battery (n = 53).

For most tasks, mean accuracy scores and standard deviations are represented as a proportion of the total number of items per test. “AMR rate” represents the average syllables per second and “adequate SMR production” illustrates the number of participants who did or did not reach criterion performance (at least three repetitions of the syllable triad /pʌtʌkʌ/). Proportions of phonological errors and semantic errors were calculated relative to the total number of errors produced by each individual participant.

Task		Mean score (SD) proportion accuracy, unless otherwise noted
IS Report		.85 (.21)
Spoken Naming		.57 (.32)
Error types during Spoken Naming (proportion of total errors)	Phonological errors	.36 (.26)
	Semantic errors	.15 (.17)
Rhyme judgment	Picture-based	.69 (.15)
	Auditory	.84 (.14)
First letter identification	Picture-based	.70 (.29)
	Auditory	.80 (.30)
Syllable count	Picture-based	.68 (.25)
	Auditory	.80 (.23)
Pseudoword repetition		.49 (.31)
Pseudoword oral reading		.35 (.36)
Motor speech evaluation	AMR rate (syllables/second)	3.42 (2.45)
	Adequate SMR production (Y:N) (at least three syllable triads)	25:28
	Segmental errors	.41 (.29)
	Ambiguous/distorted consonants	.09 (.14)

Since we planned to use difference scores for the three matched auditory/picture-based tasks in the subsequent factor analysis, we examined the relationships between IS Report and the picture-based version of each paired task: first letter identification (r = .545, p < .0001), syllable counting (r = .588, p < .0001), and rhyme judgment tasks (r = .369, p < .01). These three relationships remained significant when controlling for the auditory version of each task using partial correlations: first letter identification (r = .563, p < .0001), syllable counting (r = .624, p < .0001), and rhyme judgment (r = .364; p < .01).

3.2. Factor analysis and relationship between factor scores, IS Report, and Spoken Naming

The principal components factor analysis yielded two factors, together accounting for 60.4% of the variance in patient performance (Figure 1). The tasks requiring speech output (reading, repetition, and motor speech production) loaded primarily onto Factor 1. The retrieval measures (Picture-based – Auditory task difference scores) loaded primarily onto Factor 2.

Figure 1. Results of factor analysis.

Principal components analysis using Varimax rotation and Kaiser normalization to generate orthogonal factors. The two emerging factors relate to our measures of output (factor 1, in blue) and lexical retrieval (factor 2, in purple). Cell color is graded according to the strength of the factor loading with the strongest loadings in dark shades and weakest loadings in white. Each of the difference scores in the bottom three rows was intended to isolate phonological retrieval from other task demands and was calculated by subtracting the auditory task score from the picture-based task score.

These factor scores (Factor 1 and Factor 2) were then utilized in a set of bivariate correlations to assess the relationship of IS Report and Spoken Naming to retrieval and output (Table 2). Our hypotheses suggest that the rate at which people report successful IS should relate to individual differences in phonological retrieval ability and not to differences in output processing. Accordingly, we predicted that IS Report would relate to Factor 2 (retrieval) but not Factor 1 (output), whereas scores on Spoken Naming should relate to both retrieval and output (i.e., both factor scores). As predicted, IS Report related only to Factor 2 (retrieval); the relationship between IS Report and Factor 1 (output) is essentially null (r = .094, uncorrected p = .514). In contrast, Spoken Naming related to both Factor 1 and Factor 2. Additionally, we found a positive correlation between IS Report and Spoken Naming, consistent with both tasks relying on shared processes.

Table 2. Correlational analysis comparing logit-transformed IS Report, Spoken Naming, and factor scores.

Factor scores were generated in the context of the principal components analysis described above and presented in Figure 1. All values represent Pearson’s r;

	Spoken Naming	Factor 1	Factor 2
IS Report (logit transform)	.565*	.094	.550*
Spoken Naming	---	.602*	.394*

^*denotes statistical significance at a Bonferroni-corrected p < .05.

3.3. Relationship between IS Report, Spoken Naming, and phonological vs. semantic errors

We then examined correlations between both IS Report and Spoken Naming and naming error types. If IS judgments are a reflection of partial or complete phonological retrieval, participants who report high rates of successful IS should be likely to make phonological errors during a confrontation naming test, since phonological errors typically result from failure at the level of phonological representations or in post-lexical output processing (Nickels & Howard, 1995; Romani, Olson, Semenza, & Granà, 2002; Schwartz, Wilshire, Gagnon, & Polansky, 2004; Wilshire, 2002). As predicted, we found that IS Report related to phonological errors only (Table 3). In contrast, we found a significant, positive relationship between overall naming accuracy and semantic errors and a trending relationship between naming accuracy and phonological errors (r = .357, p = .009, Bonferroni corrected p = .08).

Table 3. Correlational analysis comparing IS Report and Spoken Naming to naming error types.

All values represent Pearson’s r;

	Phonological Errors	Semantic Errors
IS Report (logit transform)	.438*	.249
Spoken Naming	.357	.508*

^*denotes statistical significance at a Bonferroni-corrected threshold of p < .05.

3.4. The value of IS Report in predicting spoken naming ability

Finally, we performed a hierarchical regression analysis with Spoken Naming as the dependent variable to determine if IS Report provides additional explanatory power for naming ability beyond that provided by objective tests of retrieval and output processing. First, the two factor scores (retrieval and production) were entered as predictors of Spoken Naming scores. The overall model was significant (F(2,50) = 26.93, adjusted R² = .499, p < .001), and Factors 1 and 2 were both significant predictors (Factor 1 standardized beta = .602, p < .001; Factor 2 standardized beta = .394, p < .001). When IS Report was added as an additional predictor, the model fit improved significantly (Overall model F(3,49) = 29.45, adjusted R² = .621, p < .001; Model Change F(1,49) = 17.12, R² change = .125, p < .001), and only IS Report and Factor 1 (production) were significant predictors (IS Report standardized beta = .425, p < .001; Factor 1 standardized beta = .563, p < .001; Factor 2 standardized beta = .160, p = .124).

4. Discussion

The goal of this study was to clarify individual differences in the self-reported experience of successful inner speech (successful IS) in aphasia. Building on the findings from prior work (Fama et al., 2019, 2017), we examined the relationship between the subjective experience of IS and task-level performance on objective measures of different aspects of the mental process of naming. We found that the frequency of successful IS reported by an individual relates to their phonological retrieval ability and not to their speech production ability. In contrast, an individual’s success at naming aloud relates to both their phonological retrieval and speech production abilities. These findings strongly support our hypothesis that the subjective experience of IS, at least in the context examined here, relates to word retrieval and not to speech output processing.

4.1. The relationship of our findings to processing models of naming and self-monitoring

Consistent with our prior findings showing that self-reported successful IS at the item level reflects phonological retrieval (Fama et al., 2019; Hayward, 2016), the current findings clearly link the experience of IS to phonological retrieval and not to post-lexical speech production processes. It remains unclear if self-reported successful IS arises from successful retrieval of a phonological form or retrieval of a form close enough to the target to be perceived as matching it. The tasks used to objectively measure phonological retrieval only require limited phonological knowledge of the word for accurate performance. Taken as a set, accurate performance of the three phonological retrieval tasks requires knowledge of the beginning, end, and syllable number for the stimulus words, but even together they still do not require the full lexical phonological form and some elements of phonology, such as stress patterns, were not examined. The relationship between successful IS and the word retrieval factor score thus suggests that the report of successful IS reflects preserved phonological knowledge in our participants, but it does not necessarily show that successful IS reflects retrieval of the complete phonological form. Interestingly, the overall proportion accuracy was strikingly similar for all three picture-based tasks (rhyme judgment, first letter identification, and syllable counting), even though the tasks differ in the number of response options and hence in the scores that correspond with chance level performance. This suggests that participants as a group performed closer to chance (i.e., worse) on rhyme judgment than on the other two tasks, which may in turn indicate that participants’ inner speech is more likely to include the first letter and the general shape of the word than the phonology of the rime.

The results of the analyses relating IS to error types are also consistent with a relationship between successful IS and retrieval of either complete phonological forms or partially specified forms. In previous literature, there is widespread agreement that phonological errors occur either during lexical phonological retrieval or during post-lexical output processing (Nickels & Howard, 1995; Romani et al., 2002; Schwartz et al., 2004; Wilshire, 2002). The positive relationship we found between IS Report and phonological errors, therefore, supports a theory in which self-reported successful IS represents at least partial retrieval of the lexical phonological form. When naming fails in association with successful IS, it is due either to retrieval of an incompletely specified phonological form or due to errors in post-retrieval output processes after successful phonological retrieval. Individuals with apraxia of speech, a disorder of motor planning and programming, are one population that is known to have impairments in late stages of output processing, just prior to spoken output, and thus may be particularly susceptible to overt anomia after successful word retrieval. One might argue that, despite our findings, IS could still require an upstream motor component that is dissociable from spoken output. However, even in apraxia of speech, the impairment is understood to affect planning and programming stages of speech production that should impact both internal modeling and overt production of speech. Therefore, the fact that participants who exhibit some characteristics of apraxia of speech (e.g., ambiguous/distorted consonant production) can experience successful IS despite overt anomia provides additional support for the conclusion that IS does not require output processing.

Given that tip-of-the-tongue (ToT) sensations are associated with partial phonological access to words (Burke, MacKay, Worthley, & Wade, 1991; Dell, Schwartz, Martin, Saffran, & Gagnon, 1997; James & Burke, 2000; Levelt, Roelofs, & Meyer, 1999), the differences between successful IS and ToT may help to clarify the question of how much phonological information is needed to give rise to a sense of IS. The two experiences are categorically different in that with ToT, the individual by definition has not retrieved the word, whereas in describing successful IS, individuals report hearing or saying the correct word internally. Previously, we showed that IS during anomia (i.e., successful IS along with a failure of spoken naming) was related to a feeling of ToT but that each experience was associated with damage to different frontal brain regions and only anomia following successful IS (not anomia associated with ToT) was related to deficits in phonological output processing (Fama et al., 2017). This suggests that, at a minimum, successful IS reflects access to a substantially more complete form than is associated with ToT.

The level of processing at which IS arises and is available to conscious monitoring remains a matter of debate in the literature, not only in the context of aphasia but in healthy language users more generally. In naming models proposed and refined by Levelt and colleagues over the past several decades, self-monitoring of internal word forms can occur at various processing levels ranging from fully specified, pre-speech articulatory gestures (Levelt, 1983, 1989) to more abstract phonological forms (Levelt et al., 1999; Wheeldon & Levelt, 1995). In another theory of internal self-monitoring, a domain-general monitoring system is able to detect response conflict at all levels of word retrieval and production in order to identify errors (Nozari, Arnold, & Thompson-Schill, 2014). Although our work focuses on individuals with aphasia, our findings contribute to this broader scientific issue by suggesting that IS can be successfully monitored and reported on the basis of word retrieval, without a component of pre-articulatory output processing. Recent theories suggesting that motor-sensory modeling pathways are the primary substrate for internal self-monitoring (Tian & Poeppel, 2013, 2015) are at odds with our findings, which suggests that such theories cannot fully explain the process of IS self-monitoring across all individuals and all task contexts.

Beyond the level of processing at which IS is monitored, there are several other questions about the nature of self-monitoring of IS that are relevant to the broader literature on the experience in healthy language users. Further studies are necessary to determine the precise nature of the information being monitored within lexical retrieval, i.e., whether IS judgments are based on auditory images, abstract amodal phonology, or abstract word-forms (lemmas), and whether the basis of IS judgments varies across individuals and/or across tasks. Indeed, previous studies have suggested that IS is flexible and can evoke articulatory processing during certain task contexts (Oppenheim & Dell, 2010; Sokolov, 1972), so future studies should examine IS as we have here, but in other task contexts, such as silent reading or inner spelling.

4.2. Interpretation of our findings in the context of prior aphasia research

A few recent studies have investigated individual differences in IS among individuals with aphasia, but they use a definition of IS that differs significantly from our own: “the ability to create an internal representation of the auditory word form, and to apply computations or manipulations to this representation” (p. 1, Stark, Geva, & Warburton, 2017). In such studies, silent rhyme and homophone judgments on written words (Geva et al., 2011; Stark et al., 2017) or pictures (Langland-Hassan et al., 2015) are used as the measure of IS ability, emphasizing the role of IS in performing mental operations in one’s head. The mental processes that are being measured are thus not directly comparable to the form of IS we have examined in our own studies, which is a different sense of IS that involves the retrieval of words from semantic representations. Nonetheless, because they are the only prior studies specifically examining individual differences in IS among individuals with aphasia, their findings warrant consideration here.

One of these previous studies used a silent, picture-based rhyme judgment task similar to our own as a proxy for IS ability (Langland-Hassan et al., 2015). They found that participants with aphasia performed near floor on the silent task, but performed better on a 10-item auditory rhyme judgment task. Performance on the silent rhyme judgment task did not correlate with confrontation naming or generative naming ability, so the authors concluded that inner speech is often impaired across individuals with aphasia, irrespective of overt speech ability. These results are somewhat hard to interpret given the small sample size (n = 11). Indeed, in our sample there was a strong positive correlation between picture-based rhyme judgment and Spoken Naming (r = .70, p < .001) with no outliers from this relationship, suggesting that perhaps this prior study sampled a different population or was simply underpowered.

In other prior studies where IS is defined as the ability to perform mental manipulations on internal representations, it is often preserved relative to overt speech (measured by oral reading) specifically in individuals with motor speech impairments or deficits in translating from the phonological to the articulatory code (Geva et al., 2011). Such findings are consistent with the data from our own pilot work, which showed that successful IS in the context of anomia (i.e., successful IS along with a failure of spoken naming) relates to phonological output processing and to damage to sensorimotor brain regions supporting speech output (Fama et al., 2017). A more recent study (Stark et al., 2017) found relationships between IS (measured using silent rhyme and homophone judgments on written words) and spoken naming and written fluency only in people with impaired overt speech. The authors suggest a relationship between IS and working memory, error monitoring, and pre-articulation when overt speech is impaired. It is difficult to compare these results to our own because we define IS differently; in effect our studies and these are measuring different processes, although both are referred to as IS.

We made no specific predictions about which error type(s) would be associated with overall spoken naming ability, but one incidental finding of note in our study, not directly related to the experience of IS, was a positive relationship between overall Spoken Naming accuracy and the prevalence of semantic errors. This finding, although not specifically relevant to our primary research questions, is of interest because it contrasts with prior studies suggesting that there is no relationship between severity of anomia and semantic errors or that semantic errors are made with equal frequency across aphasia subtypes (Dell et al., 1997; Goodglass & Wingfield, 1997; Schwartz & Brecher, 2000). There are several reasons why this relationship, which may seem counterintuitive, may be reasonable in the context of our particular analysis. We examined errors as a proportion of the total number of errors made by each individual participant, so this finding does not indicate that individuals who are good at naming produce a higher number of semantic errors, but rather that a higher proportion of the errors that they do make are semantically related to the target. Participants with relatively mild aphasia may have tended toward semantic errors while overtly using circumlocution in search of the target word (Dell et al., 1997) or may have offered a close semantic substitute in cases where they were unable to retrieve the primary target (e.g., “bed” for target crib or “cup” for target glass). Importantly, the relationship between naming scores and semantic errors does not indicate that all participants with mild naming deficits have semantic impairments, as previous literature suggests that semantic errors in production (e.g., picture naming or oral reading) can arise during post-semantic stages of processing (Caramazza & Hillis, 1990; Dell et al., 1997; Rapp & Goldrick, 2000).

4.3. Clinical implications

Acknowledging the utility of self-reported successful IS in aphasia has important implications for clinical practice, in that a speech-language pathologist planning anomia treatment could potentially use the reported status of IS for individual items (successful or unsuccessful) to select treatment items or to match treatment approaches to specific items. The current findings reinforce the notion that overall levels of self-reported successful IS may be informative as to the main cause of anomia in a given patient, since our subjective IS measure related to overall measures of word retrieval. These data suggest that word retrieval is relatively intact in individuals who report frequent successful IS.

Including self-reports of IS along with more traditional objective speech and language measures improved predictions of spoken naming ability, which demonstrates the potential practical value of using IS judgments to improve anomia diagnosis. In a clinical setting, a speech-language pathologist could utilize a silent-picture naming test, in addition to a spoken naming test, to gather more information about the status of word retrieval. Overall levels of self- reported successful IS could be combined with information gathered via other common approaches to anomia diagnosis such as error patterns (Laine & Martin, 2006), response to cueing (Howard & Gatehouse, 2006; Wambaugh et al., 2001), and overall pattern of language ability (Goodglass & Wingfield, 1997; Howard & Gatehouse, 2006; Laine & Martin, 2006), in order to provide a more comprehensive and accurate assessment of the locus of word-finding impairment. The current findings show that successful IS may provide diagnostic information that is complementary to error patterns and objective tests of speech and language ability. Future research could examine relationships between self-reported successful IS and response to phonological cueing, which would help to better understand participants’ level of phonological knowledge about words that they cannot say out loud.

4.4. Consideration of potential biases in self-reported inner speech

Importantly, the significant relationship identified between the subjective experience ofsuccessful IS and word retrieval suggests more generally that IS self-reports are often valid and meaningful. We found that participants, on average, reported successful IS for a greater number of items than they were able to produce accurately during spoken naming. We expected this to be the case, since it is likely that some people with aphasia have relatively intact word retrieval alongside production deficits, which would cause errors during spoken naming despite a judgment of successful IS. Although this explanation aligns with our hypotheses and is supported by analyses reported elsewhere (Fama et al., 2019), we also cannot rule out the possibility that some (or all) participants may be over-reporting the success of their IS to some degree.

In general, given the group-level analysis performed here, we cannot comment on individual variability in self-monitoring for IS in the context of these particular findings. In prior work (Fama et al., 2019, 2017), we have considered the potential impact of self-monitoring on the accuracy of self-reported successful IS in detail. One specific prediction about individual reliability would be that individuals whose pattern of impairments are consistent with Wernicke’s aphasia would likely be unreliable in IS self-monitoring, as these individuals show poor monitoring for their overt speech (J. Marshall, Robson, Pring, & Chiat, 1998; R. C. Marshall, Neuburger, & Phillips, 1994). We excluded participants with severe comprehension deficits from our studies since they might be unable to understand the task instructions, so further research would be required in order to assess reliability of self-reported IS in this particular population. Overall, we found significant statistical relationships between our subjective measure of IS and several predicted objective measures, so it is likely that the majority of our participants were largely accurate in reporting the success of their IS.

Our measure of successful IS is entirely subjective in nature so is potentially subject to individual bias in participants beyond those with self-monitoring impairments. For instance, there may be a third variable that happens to covary along with both self-reported IS and with phonological retrieval, e.g., level of confidence about task performance or individual variability in the threshold for judging accuracy. Within the context of our silent picture-naming task, participants were allowed only a binary response (yes/no) for reporting the success of inner speech. In the future, a more nuanced version of the task might offer participants the opportunity to provide more detail about what a “no” response means, i.e., have they retrieved a word related in meaning, or in sound, or no word at all?

Additionally, it may be beneficial in future studies to add an accuracy judgment for spoken naming tasks in order to evaluate self-monitoring of overt speech, which could in turn help clarify participants’ likelihood of over-reporting successful IS. An essential caveat, however, is that assessing accuracy of self-reported IS is extremely difficult, if not impossible. For instance, even if we could demonstrate our participants to be highly accurate in judging their spoken responses, there is no guarantee that they are judging their IS with similar accuracy. Many models of speech self-monitoring suggest different error monitoring mechanisms operating at different levels of word retrieval and production (e.g., Postma & Noordanus, 1996), so it may be possible for participants to be accurate in judging their IS but not their spoken output, or vice versa. Because of this inherent difficulty in assessing accuracy of self-reported IS, the focus of our research is not on the accuracy of these reports, but whether they provide useful information about mental processes or clinical outcomes. For instance, we have previously demonstrated that items reported as successful IS prior to anomia therapy are relearned faster than items reported as unsuccessful IS (Hayward et al., 2015). We have also demonstrated that patients have far more phonological knowledge about items that they report naming successfully internally than those they report being unable to name internally (Fama et al., 2019). Thus, although we acknowledge that there may be some over-reporting of successful IS, this does not negate the utility of these reports or our findings that they are strongly related to phonological retrieval ability and not to speech production ability.

5. Conclusions

This study provides robust evidence that the experience of successful IS in aphasia is most common among individuals who have relatively spared lexical/phonological retrieval. Participants’ speech output processing abilities support their spoken naming ability, but do not relate to their experience of successful IS. Along with the spoken naming error patterns associated with successful IS, these findings demonstrate that the subjective experience of successful IS relates to phonological retrieval and that speech output processes are not a necessary component of IS. Further, adding IS reports to objective measures improves predictions of spoken naming ability, demonstrating the potential clinical value of this subjective measure. We suggest that self-reported successful IS can be an informative addition to the diagnosis of anomia in the context of processing models of naming, which in turn may inform clinical decision-making regarding anomia treatment.

Highlights.

• Individuals with aphasia often report that they can say words in their heads, despite overt anomia

• Silent picture-naming can be used as a measure of self-reported inner speech (IS) ability

• Self-reported IS relates to performance on phonological retrieval tasks and not output processing tasks

• Self-reported IS relates to higher proportion of phonological errors on spoken naming tasks